A pack is data: it names deterministic checks that already exist in this repository and points at the committed results file that contains them. It cannot add a provider, a credential or a spend. The validator refuses a pack that claims a pass the results file does not contain.
Evidence is read from benchmarks/results/latest.json, generated 2026-11-15 13:30 UTC from source951b313. A recorded pack is a dated observation, not a live claim about this exact checkout.
CF-COMPOSE-01 · cf-compose-01.json
Cloudflare composition: local evidence foundation
Not run
Reuses all eleven sandbox save/reopen and blueprint round-trip checks, then exercises the real read-only provider operations view with fictional scope, coverage, dependency and cost observations. It establishes local contracts, not live Cloudflare capability.
Not run. This pack has no passing evidence in the committed file, so it is not shown as a result. No committed baseline is attached to this pack. Retain each source-pinned local output separately; fixture results do not establish completed Cloudflare journeys.
A local pass covers the named source functions and fictional observations only; it is not a browser, deployed application or live Cloudflare check.
Provider operations are derived from supplied observations. No account is contacted, and no actual resource permissions, provisioning or preview isolation are established.
Only observed service/D1/R2/KV bindings are represented. Queues, Workflows, external clients and unread Worker settings remain outside the known dependency graph.
This pack does not measure model quality, production latency, upload throughput, provider charges, data restoration or completion of the Cloudflare wishlist.
The zero-spend requirement prohibits provider execution here; it does not turn unknown cloud resource charges into zero costs or grant retirement authority.
Run this optional pack explicitly through the existing runner; its separate directory does not add a blanket check to routine releases.
FLOW-01 · flow-01.json
Sandbox flow: save, reopen and export a blueprint
Recorded evidence
A provider-free pack over the public sandbox's deterministic store, moderation and blueprint checks. It proves the same edits survive a save and reopen, a tampered save is refused, a hidden note stays hidden after restore, and a restored sandbox still exports a valid blueprint.
A save with no contributions and a save with version 2 were both refused; each recovered to the initial sandbox instead of half-applying.
✓
restore-keeps-moderation
Version 1 restored the saved name while note-1 stayed hidden and note-2 stayed pending; appearance and moderation travel separately.
✓
export-still-reads-restored-state
Export after restore produced preset “composition” with 2 links.
blueprint-roundtrip
✓
export-succeeds
Exported version 2 preset “composition” with 3 links.
✓
fields-survive
Name, introduction, theme, layout, both enabled blocks and 3 links survived export → JSON → re-import.
Evidence recorded 2026-11-15 13:30 UTC from source951b31384a6abb85f947fbc46c87a3ee95363c63. The pack checks the same ids in the committed file.
The pack runs only the deterministic checks named here. It does not measure model quality, latency or cost.
A pass means these in-repo helpers held on the committed source; it is not a provider, deployment or browser check.
The evidence is a dated run. A later source revision can make the result stale, and the page says so rather than assuming it still holds.
OBS-01 · obs-01.json
Offline observability fixture
Not run
Check the relationship between a failed request, its spans and a log. The fictional fixture includes a later recovery, an independent control and missing, duplicate and unsupported signals.
Provider free
Yes
Network
Not used
Credentials
None
Spend
$0
Checks
4 deterministic
Evidence
none yet
obs-01
·
observed-correlated-failure
Not present in the committed results file.
·
observed-independent-recovery
Not present in the committed results file.
·
duplicate-or-missing-unproven
Not present in the committed results file.
·
unsupported-signal-unproven
Not present in the committed results file.
Not run. This pack has no passing evidence in the committed file, so it is not shown as a result. The catalogue has no committed run receipt. Run the offline fixture and open its JSON result below.
These are fictional request, log and trace records. Their timings come from the fixture timestamps.
The checks run locally with zero provider use or spend. Live telemetry delivery, incident diagnosis and remediation require separate evidence.
Can App Composer build App Composer?
The challenge is to rebuild and operate App Composer and its supported apps in an independently owned Cloudflare account, then repeat the process from the rebuilt instance. If a creator can reproduce this system from documented inputs, they can use the same method to understand, own and evolve their own projects.
Pin A. Freeze a baseline and publish the permitted blueprints and contracts.
Build B. Use a supported, authorised agent in a separate account with its own credentials. Use only the published inputs and record any assistance.
Repeat from B. Create and evolve an app using B, then use B to construct C. Compare both rebuilds with the frozen parity checklist and record any gaps.
First milestone · A → B
Deploy a functional Composer slice in a separately owned account and exercise real auth, data, media and operations. B → C starts after that evidence is available. Routine releases continue independently of this challenge.
Before a run
Exact source versions, reusable block manifests, compatibility and licence/rights checks
Isolated independent Cloudflare account/credentials explicitly authorised by that owner
Freeze functional parity assertions, exclusions and cost limits before any attempt
The runner validates the pack, resolves the named modules, runs only those checks and prints the same result shape as a full run. Without --suite, the existing .mjs suite discovery is unchanged.
Inspect a local run result
Open a runner result here to inspect it in this tab.